RNA Modification Detection using Nanopore Direct RNA Sequencing via improved Transformer

RNA modifications are a common occurrence in transcriptome and play a crucial role in various biological processes. Nanopore direct RNA sequencing (DRS) provides raw current signal readings, which carry information of modifications. Supervised machine learning methods using DRS are advantageous for RNA modification detection. However, existing methods for RNA modification detection do not adequately capture sequential signal features within and between reads.Here, we represent NSWord, an improved transformer model with three types of self-attention blocks that integrates the transcript sequence and its signal reads to produce a comprehensive site-level prediction. NSWord outperforms existing deep learning methods, particularly in terms of precision of top-scoring predictions. Additionally, we investigate the impact of limiting the length or number of signal reads and explore the role of transcript sequence in modification prediction.